Toward Unsupervised 3D Reconstruction From Unorganized Multi-view Images

Simultaneous estimation of camera motion and structure of static scene using uncali-brated images from multiple views is a common task in computer vision and of interest formany applications. The theory of multi-view geometry has been intensively studied in thelast two decades and nowadays is subject for textbooks.In this thesis, the challenging problem of unsupervised metric reconstruction from alarge set of unorganized still images is tackled. The difficulties arise from two main aspects:(1) For ordered image set like video sequences, the reconstruction is straight-forward with ahierarchical process. For unordered image set without any prior information, the desidera-tum is to find a strategy that can achieve successful reconstructions; the state-of-the-art canonly effectively deal with small number of images. (2) For most of the existing systems, alot of professional human-computer interaction is needed during the reconstruction process.To achieve completely unsupervised multi-view reconstruction, our system must faces newchallenges from the lack of robustness.A series of new algorithms are presented to improve the overall robustness and per-formance. Finally, a novel reconstruction strategy based on graph theory, which has nodependence on the order of the input images, is proposed.RANSAC(RANdom SAmple and Consensus) is the standard choice for epipolar geom-etry estimation. We addressed some problems of RANSAC posed from both practical andtheoretical standpoints and two new algorithms is presented: RANSAC with adaptive Tc,dtest and the GMSAC(Gaussian Mixture based SAmple and Consensus). Adaptive Tc,d testis an extension of the original RANSAC, which can achieve user independent RANSAC ac-celeration. GMSAC adopts the random sample strategy and maximization likelihood theory,and gaussian mixture is used to model different types of outliers respectively.For projective reconstruction, a linear algorithm with reprojection error minimizationis derived. Our implementation is based on the weighted alternated least square between thestructure, motion and projective depth. Our method can be used as the booting or compositealgorithm for the standard projective bundle adjustment, which could improve the overallreconstruction quality and performance.For self-calibration and metric upgrade, a method based on the two-way quasi-affinereconstruction is introduced. Before upgrade to the metric, firstly upgrade to the quasi-affine could increase the chance to reach the global optimal, and so improve the final metric quality.System test was run on several different image sets, including outdoor and indoorimages. We show experimentally that our system can achieve high quality metric recon-struction on complex unorganized image set without any human-computer interaction.